corr {lares}

Correlation table

Description

This function correlates a whole dataframe, running one hot smart encoding (ohse) to transform non-numerical features. Note that it will automatically suppress columns with less than 3 non missing values and warn the user.

Usage

corr(
  df,
  method = "pearson",
  use = "pairwise.complete.obs",
  pvalue = FALSE,
  padjust = NULL,
  half = FALSE,
  dec = 6,
  ignore = NULL,
  dummy = TRUE,
  redundant = NULL,
  logs = FALSE,
  limit = 10,
  top = NA,
  ...
)

Arguments

df

Dataframe. It doesn't matter if it's got non-numerical columns: they will be filtered.

method

Character. Any of: c("pearson", "kendall", "spearman").

use

Character. Method for computing covariances in the presence of missing values. Check stats::cor for options.

pvalue

Boolean. Returns a list, with correlations and statistical significance (p-value) for each value.

padjust

Character. NULL to skip or any of p.adjust.methods to calculate adjust p-values for multiple comparisons using p.adjust().

half

Boolean. Return only half of the matrix? The redundant symmetrical correlations will be NA.

dec

Integer. Number of decimals to round correlations and p-values.

ignore

Vector or character. Which column should be ignored?

dummy

Boolean. Should One Hot (Smart) Encoding (ohse()) be applied to categorical columns?

redundant

Boolean. Should we keep redundant columns? i.e. If the column only has two different values, should we keep both new columns? Is set to NULL, only binary variables will dump redundant columns.

logs

Boolean. Calculate log(x)+1 for numerical columns?

limit

Integer. Limit one hot encoding to the n most frequent values of each column. Set to NA to ignore argument.

top

Integer. Select top N most relevant variables? Filtered and sorted by mean of each variable's correlations.

...

Additional parameters passed to ohse, corr, and/or cor.test.

Value

data.frame. Squared dimensions (N x N) to match every correlation between every df data.frame column/variable. Notice that when using ohse() you may get more dimensions.

See Also

Other Calculus: dist2d(), model_metrics(), quants()

Other Correlations: corr_cross(), corr_var()

Examples

data(dft) # Titanic dataset
df <- dft[, 2:5]

# Correlation matrix (without redundancy)
corr(df, half = TRUE)

# Ignore specific column
corr(df, ignore = "Pclass")

# Calculate p-values as well
corr(df, pvalue = TRUE, limit = 1)

# Test when no more than 2 non-missing values
df$trash <- c(1, rep(NA, nrow(df) - 1))
# and another method...
corr(df, method = "spearman")